基于遗传算法减少语音助手虚假唤醒词的设计与实现

Design and Implementation of Reducing False Wake-up Words of Voice Assistant Based on Genetic Algorithm

在物联网领域,语音助手已经成为操作智能音箱、智能手机,甚至汽车的重要方式之一。为了节省电力和保护用户隐私,语音助手只有在检测到唤醒词时才会向云端发送命令。然而,语音助手很容易受到虚假唤醒的影响从而虚假唤醒。基于此,本文从3个方面对虚假唤醒现象进行了研究。首先,设计了一个模糊词生成器,能自动有效地产生模糊词,而不是通过大量的音频材料搜索模糊词。其次,为了解释虚假唤醒现象背后的原因,构建了一个可解释的基于树的决策模型,该模型揭示了唤醒词检测器错误接受模糊词的语音特征。最后,提出了减轻虚假唤醒影响的补救措施。结果表明,强化后的模型不仅能拒绝模糊词的虚假唤醒,而且在原始训练数据集上取得了更好的整体性能。

In the Internet of Things field,voice assistants have become one of the most important ways to operate smart acoustic enclosure,smart phones and even cars.To save electric power and protect user privacy,voice assistants only send commands to the cloud terminal when a wake-up word is detected.However,voice assistants are susceptible to false wakeups.Based on this condition,this paper investigates the false wake-up phenomenon in three aspects.First,a fuzzy word generator is designed to automatically and efficiently generate fuzzy words,instead of searching for fuzzy words through a large amount of audio materials.Secondly,to explain the reasons for the false wake-up phenomenon,an interpretable treebased decision model is constructed to reveal the speech characteristics of the wake-up word detector that incorrectly accepts fuzzy words.Finally,remedial measures are proposed to mitigate the effects of false work-ups.The results show that,the enhanced model can not only reject false work-ups of fuzzy words,but also achieve better overall performance on the original training dataset.